sound recognition of dc machine with application of fft and fuzzy logic
Transkrypt
sound recognition of dc machine with application of fft and fuzzy logic
Nr 62 Prace Naukowe Instytutu Maszyn, Napędów i Pomiarów Elektrycznych Politechniki Wrocławskiej Nr 62 Studia i Materiały Nr 28 2008 feature extraction, fuzzy logic, identification, recognition, sound. Adam GŁOWACZ*, Witold GŁOWACZ* SOUND RECOGNITION OF DC MACHINE WITH APPLICATION OF FFT AND FUZZY LOGIC A new approach to determination of similarity of dc machine sound is presented. This approach is based on FFT and fuzzy logic. Investigations of the sound recognition were carried out for direct current machine. The results of sound recognition are included in this paper. 1. INTRODUCTION At present there are many methods of sound recognition [2], [5], [14]. Most of them are based on data processing [4], [8], [11], [16]. The aim of this paper is analysis of a system which enables sound recognition. To protect electrical machine the shorted coils are detected in rotor winding. Sound recognition application contains preliminary data processing, feature extraction and classification algorithm. It makes possible to identify sounds. One of the tasks was to use fuzzy logic. Fuzzy logic functions were applied as a classifier [3], [6], [12], [17]. Investigations were carried out for direct current machine because it produces characteristic sounds. It can notice that sound of faultless dc machine is different from sound of faulty dc machine [1], [7], [9], [10], [13], [15]. For recognition aim the mechanism of early detection of damages in dc machine was created. 2. SOUND RECOGNITION PROCESS Sound recognition process contains fuzzy logic set creation process (Fig. 1) and identification process (Fig. 2). At the beginning of fuzzy logic set creation process the __________ * University of Science and Technology, Faculty of Electrical Engineering, Automatics, Computer Science and Electronics, al. Mickiewicza 30, 30-059 Kraków, Poland, e-mail: [email protected] 499 signals are sampled and normalized. Afterwards data are converted through the Hamming window. Next data are converted into a frequency band through the fast Fourier transform. The fast Fourier transform creates feature vectors. Fuzzy logic set creation process and identification process are based on the same signal processing algorithms. The difference between them is a sequence of execution. All feature vectors are averaged in fuzzy logic set creation process. Two averaged feature vectors are created. Afterwards it is converted into fuzzy logic set. Fuzzy logic set creation process contains following steps: sampling, quantization, normalization, filtration, windowing, feature extraction (two averaged feature vectors) and fuzzy logic set formation. Fig. 1. Fuzzy logic set creation process Classification is used in the identification process. It is based on fuzzy logic. There was applied fuzzy logic as a classifier. To obtain results of recognition, it compares feature vector of new sample with averaged feature vector with the help of fuzzy logic functions. Identification process contains following steps: recording of acoustic signal, sound track division, sampling, quantization, normalization, filtration, windowing, feature extraction, classification. 500 Fig. 2. Identification process 2.1. ACOUSTIC SIGNAL RECORDING The sound card with analogue-digital converter is able to record, process and replay sound. The recording of the acoustic signal is the first part of identification process. Acoustic signal is converted into digital data (wave format) by the microphone and the sound card. This wave file contains following parameters: sampling frequency is 16000 Hz, number of bits is 16, number of channels is 1 (mono). 2.2. SOUND TRACK DIVISION Application divides sound track into sound fragments. It divides data. Next it creates new wave header. Afterwards new wave header is copied. Then new wave header is added to each chunk of data. New wave files are obtained. These files are used in the identification process. There are following advantages of such solution: precise determination of sound appearing, precise sound identification, application does not have to allocate as much memory in identification process. 501 2.3. SAMPLING Sampling is a technique to convert an analog signal into a digital signal. It periodically samples an input signal and transforms into a sequence of intensity values. Sampling frequency is basic parameter. Sampling frequency is 16000 Hz in sound recognition application (Fig. 3). Fig. 3. Sound of dc machine with shorted coils for five seconds before normalization 2.4. QUANTIZATION Quantization is a technique to round intensity values to a quantum so that they can be represented by a finite precision. Precision of sample values is specific to number of bits. Common applied number of bits is 8 or 16. Sound recognition application uses 16 bits because it gives better precision. There is a choice of number of bits depending on quantity of input data and calculations speed in sound recognition process. The compromise is important to obtain good results in short time 2.5. NORMALIZATION In sound recognition application, the normalization is the process of changing of the amplitude of an audio signal. There is a possibility that some sounds aren’t recorded at the same level. It is essential to normalize the amplitude of each sample in order to ensure, that feature vectors will be comparable. All samples are normalized in the range [−1.0, 1.0]. In method the amplitude maximum of the samples is found and then each sample is divided by this maximum. 502 2.6. FILTRATION Filtration is a very efficient way of removing the unwanted noise from the spectrum. The filtration is used to modify the frequency domain of the input sample. The filtration is not necessary to sound recognition. However the usage of this can improve the efficiency of the sound recognition. 2.7. WINDOWING Windowing is a technique used to shape the time portion of measurement data, to minimize edge effects that result in spectral leakage in the FFT spectrum. By using window functions, the spectral resolution of frequency domain will be increased. There are different types of window functions available, each with their own advantage. The Hamming window is used to avoid distortion of the overlapped window functions 2.8. FAST FOURIER TRANSFORM The sequence of frequency of a signal obtained by FFT becomes the basis for extracting of the frequency-domain features. It is applied instead of discrete Fourier transform because of shorter time of calculations. Obtained coefficients create feature vectors which are used in calculations. 2.9. CLASSIFICATION Difference between sounds depends on differences in ordered sequence. Classification uses feature vectors and fuzzy logic functions in the identification process. It compares different values of feature vectors. It compares feature vectors with the help of fuzzy logic functions (feature vector of investigated sample, feature vector of specific category). Fuzzy logic functions use amplitude of the sample to determine probability (Fig. 4–5). If probability of determined fuzzy function is greater than 0.5 and then function is chosen (240 functions). There are some points where probability is 0.5 and then one of fuzzy function is chosen. Category including the most correct values is the result of the identification 0 x−a 1 a2 − a1 p ( x) = a3 − x a3 − a2 0 x < a1 a1 ≤ x < a2 (1) a2 ≤ x < a3 x ≥ a3 503 Fig. 4. Frequency spectrum of sound of faultless dc machine for ten seconds after normalization with fuzzy logic functions. Fig. 5. Fuzzy logic function. 3. SOUND RECOGNITION RESULTS Investigations were carried out for sound of faultless dc machine and sound of dc machine with shorted coils. Ten five-second samples were used for fuzzy logic set creation process for each category. New samples which were unknown for the system were used in the identification process. The system should specify the state of dc machine correctly. The fuzzy logic set creation process was carried out for five-second samples. Identification process was carried out for two-second, three-second, foursecond, five-second, six-second and seven-second samples. Sound recognition efficiency in dependence on length of sample is presented in Fig. 6. Sound recognition efficiency [%] 504 120 100 80 60 40 20 0 2 3 4 5 6 Length of sample [s] 7 Sound of faultless dc machine Sound of dc machine with shorted coils Fig. 6. Sound recognition efficiency in dependence on length of sample with the low-pass filter which passes frequencies below 3062 Hz. 4. CONCLUSIONS Sound recognition application was created. It identifies category including the most correct values. The algorithms of signal processing and fuzzy logic were used in the identification process. Analysis shows the sensitivity of methods which are based on fuzzy logic algorithm in dependence on input data. Investigations show that fuzzy logic functions work for different input data. The best results were obtained for sixsecond samples. It also used the low-pass filter which passes frequencies below 3062 Hz. The sound recognition efficiency was 71.42% for faultless dc machine whereas the sound recognition efficiency was 100% for dc machine with shorted coils. Time of identification process of one sample was 0.625 s for Intel Pentium M 730 processor. REFERENCES [1] ANTAL M., ANTAL L., ZAWILAK J.: Badania eksperymentalne silnika indukcyjnego z uszkodzoną klatką wirnika,. Prace Naukowe Instytutu Maszyn, Napędów i Pomiarów Elektrycznych Politechniki Wrocławskiej; nr 59, Studia i Materiały, Oficyna Wydaw. PWroc., 2006, s. 69–77. [2] CAMPBELL J.P., Speaker Recognition: A Tutorial, Proc. IEEE, Vol. 85, Issue 9, IEEE Press, pp. 1437–1462, 1997. 505 [3] CHEN Y. C., TSAO T. C., A Description of the Dynamical Behavior of Fuzzy System, IEEE Transactions on Systems, Man, and Cybernetics, Vol. 19, No. 4, July/August 1989. [4] CHILDERS D. G., Speech Processing and Synthesis Toolboxes, Wiley, New York, 1999. [5] COUVREUR C., Environmental Sound Recognition : a Statistical Approach, PhD thesis, Faculté Polytechnique de Mons, Belgium, June 1997. [6] DUBOIS D., PRADE H., Fuzzy Sets and Systems: Theory and Applications, Academic Press, New York, 1980. [7] DUDZIKOWSKI I., PAWLACZYK L.: Maszyny prądu stałego o magnesach trwałych i ich sterowanie – stan aktualny i perspektywy rozwoju. Analiza, projektowanie i sterowanie maszyn elektrycznych. Gliwice: Wydaw. PŚl. 2001, s. 59–78. [8] GŁOWACZ A., GŁOWACZ W., Sound recognition of dc machine with application of FFT and backpropagation neuronal network, Proceedings of 12th PPEEm Symposium on Fundamental Problems of Power Electronics Electromechanics and Mechatronics, PPEEm, Wisła, 9–12 December, 2007, pp. 269–274. [9] GŁOWACZ Z., GŁOWACZ W., Mathematical Model of DC Motor for Analysis of Commutation Processes, Proceedings of 6th IEEE International Symposium on Diagnostics for Electric Machines, Power Electronics and Drives (SDEMPED), IEEE, Cracow, 6–8 September, 2007, pp. 138–141. [10] GŁOWACZ Z., ZDROJEWSKI A., Spectral Analysis of Signals of Commutator DC Motor, Proceedings of XVII International Conference on Electrical Machines (ICEM 2006), CD, No. 656, Chania, Greece, 2006, pp. 1–4. [11] GOLDHOR R. S., Recognition of Environmental Sounds, Proceedings of ICASSP Vol. 1, pp 149–152, New York, NY, USA, April 1993. [12] KOSANOVIC B. R., CHAPARRO L. F., SCLABASSI R. J., Signal Analysis In Fuzzy Information Space, Fuzzy Sets and Systems, Vol. 77, pp. 49–62, 1996. [13] KOWALSKI C. T.: Monitorowanie i diagnostyka uszkodzeń silników indukcyjnych z wykorzystaniem sieci neuronowych, Wyd. Pol. Wrocławskiej, Monografie, t.57, nr 18, Wrocław, 2005. [14] LOVEKIN J. M., YANTORNO R. E., KRISHNAMACHARI K. R., BENINCASA D.B., WENNDT S. J., Developing usable speech criteria for speaker identification, IEEE, International Conference on Acousitcs and Signal Processing, pp. 424–427, May 2001. [15] MAKOWSKI K.: Space harmonic EMFs in rotor bars of single-phase shaded pole induction motors. Harmoniczne przestrzenne SEM w prętach wirnika jednofazowych silników indukcyjnych z pomocniczym uzwojeniem zwartym.: Sixth International Symposium on Electric and Magnetic Fields. EMF 2003. Aachen, Germany, 6–9 October 2003. pp. 333–336 [16] RABINER L., JUANG B.-H., Fundamentals of Speech Recognition, Englewood Cliffs (N.J.), Prentice Hall Signal Processing Series, 1993. [17] SZABAT K., ORŁOWSKA-KOWALSKA T.: Adjustment of classical and fuzzy logic speed controllers for electrical drives with elastic joint, Proceedings of XVI International Conference on Electrical Machines (ICEM 2004), Cracow, 2004. ROZPOZNAWANIE DŹWIĘKU MASZYNY PRĄDU STAŁEGO Z ZASTOSOWANIEM FFT I LOGIKI ROZMYTEJ Przedstawiono nowe podejście do rozpoznawania dźwięków maszyny prądu stałego. Podejście to jest oparte na zastosowaniu szybkiej transformacji Fouriera i logiki rozmytej. Badania przeprowadzono dla maszyny prądu stałego. Wyniki badań potwierdzają dużą skuteczność rozpoznawania dźwięku maszyny prądu stałego.